Honte, a go-playing program using neural nets
Fredrik Andreas Dahl · 2001
The go-playing program Honte is described. It uses neural nets together with more conventional AI-methods like alpha-beta search. A neural net is trained by supervised learning to imitate local shapes made in a database of expert games. A second net is trained to estimate the safety of groups by self play using TD(l)- learning. A third net is trained to estimate territorial potential of unoccupied points, also based on self play and TD(l)-learning. Although the program has not yet reached the level of the best commercial go-programs, results are encouraging. 1 INTRODUCTION This article describes the go-playing program Honte. The name "Honte" (pronounced hon-teh) means "proper", "sound" or "correct" in Japanese. The idea behind Honte is to use neural nets together with other programming techniques, hopefully getting the best of all worlds. Not all sections of the article are relevant to machine learning as such, but for a problem as complex as go, integration of different techniques i...